Image Reconstructor (from Patches)

Python code to reconstruct images from patches with a specified patch size/stride combination. The patches are combined by averaging, making this useful when reconstructing an image from imperfect patches generated from a neural network.

This code is more flexible than the Scikit reconstruct_from_patches_2d function since it can work with any stride value.

Contents

The patch_reconstructor package contains the following functions within recon_from_patches.py:

• recon_im: Reconstructs an image from an array of patches with a specified stride and patch size. Overlapping areas are resolved by averaging.

• get_patches: Extracts overlapping patches (square) of a specified size and stride from an input image. This is a demo function used to prove the functionality of the recon_im function.

Installation

Activate your preferred Python (3.x) environment and run the following command from the repo home directory:

pip install .

Numpy is the only dependency for this package.

Usage

You can use this function directly in Python after installation by first importing it using the following statement:

from patch_reconstructor.recon_from_patches import recon_im

Then, the actual reconstruction function (recon_im) can be called with the following parameters:

recon_im(patches: np.ndarray, im_h: int, im_w: int, n_channels: int, stride: int)

patches: 4D ndarray with shape (patch_number,patch_height,patch_width,channels)
    Array containing extracted patches. If the patches contain colour information,
    channels are indexed along the last dimension: RGB patches would
    have `n_channels=3`.
im_h: int
    original height of image to be reconstructed
im_w: int
    original width of image to be reconstructed
n_channels: int
    number of channels the image has. For  RGB image, n_channels = 3
stride: int
       desired patch stride

Demo

To execute a short demo run the following commands from the repo home directory:

1. pip install -r requirements.txt (adds pillow, click and matplotlib for graphing functionalities)
2. pip install .
3. python demo.py samples/baboon.bmp

The demo will produce an image showing the patching and reconstruction process. The image, stride and patch size can be adjusted as required.

Further Development

Open to suggestions/improvements.